TL;DR
This paper introduces a statistical method to detect changes in neuronal firing rates, even when these changes are not perfectly synchronized across trials. They develop estimators for the distribution of these change points, showing that while individual change points are hard to pinpoint, their overall distribution can be reliably estimated from neural data.
What Problem Does This Paper Address?
Neuroscientists often study how neurons respond to external events by looking at their firing patterns over many trials. A common technique, the Peri-Stimulus Time Histogram (PSTH), averages activity across trials. However, this can mask important details: a smooth change in firing rate within each trial looks similar to sharp changes that happen at slightly different times across trials. This paper tackles the challenge of precisely identifying and characterizing these changes in neuronal activity, especially when the timing of the change is variable from trial to trial. They aim to distinguish between a gradual shift in firing rate and a consistent, abrupt shift with timing jitter.
Methods
The authors propose a statistical framework for detecting change points in neuronal activity. They model the firing rate of a neuron as an inhomogeneous Poisson process that changes intensity at one or more points in time. Crucially, the timing of these change points is treated as a random variable with an unknown distribution. They develop both simple and non-parametric maximum likelihood estimators (NPMLE) for this change point distribution, proving that these estimators achieve the best possible convergence rates. The paper also extends this model to handle multiple neurons and multiple change points per trial. For estimation, they employ the Expectation-Maximization (EM) algorithm, a standard technique for models with missing data, to iteratively refine parameter estimates.
Datasets
The paper uses empirical data recorded from two awake vervet monkeys performing a visual-motor task. This data consists of spike trains from neurons in the basal ganglia. They also use simulated data to test and validate their methods.
Key Insights and Ideas
- Distinguishing smooth vs. abrupt changes: The core contribution is the ability to differentiate between a gradual change in firing rate and a sharp, but variably timed, change across trials. This is crucial for understanding neural processing.
- Distribution of change points is key: While precisely locating the change point in any single trial can be very difficult and prone to error, the distribution of these change points across many trials can be estimated reliably. This is a significant finding for interpreting neural responses.
- Empirical Bayes approach: The method aligns with an empirical Bayes framework, leveraging information across trials to improve estimates, especially for the change point distribution.
- Model flexibility: The proposed model can accommodate various distributions for the change point timing, making it adaptable to different neural response patterns.
Key Results
- The proposed estimators for the change point distribution achieve optimal convergence rates, meaning they are as efficient as statistically possible.
- For realistic spike train data, trial-by-trial estimation of a change point can be misleading, but the distribution of change points can be well-estimated.
- The paper demonstrates the method's effectiveness on both simulated and real neural data, showing successful identification of change points and their distributions in various scenarios, including single and multiple change points, and single and multiple neurons.
Limitations and Future Work
- Refractory periods: The Bernoulli model used for spike trains is an approximation and doesn't fully account for neuronal refractory periods, though it's considered valid if the refractory period is short compared to the mean inter-spike time.
- Computational complexity: The EM algorithm can be slow, especially for extensions with multiple change points.
- Model appropriateness: For some cells, the change point model might not be entirely appropriate, or a larger sample size might be needed for stable estimation.
- Future work: The authors suggest exploring more complex models, such as those accounting for the duration of spikes, and further investigating the testing of change point existence versus no change, and distinguishing between sharp and gradual changes.
Summary updated